Agreed that your hypothesis makes sense given the current state of things but I do not think it scales with any sort of consistency/quality control in the utopian "no labels all self-published" world.

This sounds terrible from a user perspective and arguably promotes certain types of music while penalizing others. As a user, if I hear more than a few songs that are terrible that were forced upon me by some algorithm, I am not going to listen to that playlist/use that service anymore. If users are the only way you are generating training data (for this supposed only self-published music world), you are going to end up with the music version of clickbait to get users' attention before they hit skip. I really cannot emphasize how bad optimizing for this would be for music.

Whatever your opinion on "current gatekeepers", some of them do provide valuable filtering services that I struggle to see how an algorithm could accurately replicate without compromising what makes music magical.

I've watched some of Paul's talks as well as seeing some of his early demos at Sun Labs. The bottom line is that discovery is really hard generally and, for the most part, no one has cracked it pretty much anywhere even when there aren't competing commercial interests.